Advancing AI Safety Across Languages, Cultures and Contexts

Detailed Summary

• Natasha Crampton opened the session by emphasizing India’s linguistic richness as a fitting backdrop for a conversation on multilingual AI safety.
• She argued that multilingual and multicultural capability is not a luxury but a necessity for the broad, beneficial diffusion of AI.
• Key data points:
  • Fewer than 5 % of spoken languages are meaningfully represented on the public web.
  • English accounts for ~42 % of widely used training datasets.
• The imbalance leads to safety mechanisms calibrated on English‑centric norms, which miss:
  • Indirect phrasing, idioms, humour, and culturally specific expressions.
  • Contextual harms such as scams, manipulation, and social engineering where intent is hinted rather than explicit.
• Low‑resource languages and non‑Latin scripts have been exploited by adversarial actors to bypass safeguards that were never tested in those contexts.

2. Microsoft’s Multilingual Safety Initiatives

2.1 AI Luminate Benchmark (Multilingual, Multicultural, Multimodal)

• Peter Mattson (MLCommons) was referenced as a lead on the AI Luminate benchmark, which is being expanded to cover many languages, cultures, and modalities (text, vision, audio).

2.2 Samakisha – Community‑Centred Evaluation

• Sunayana Sitaram (Microsoft Research, India) and her colleague Kalika (also present) described Samakisha, a community‑driven framework that evaluates model behaviour in real‑world contexts by involving local users and experts.

2.3 Project Gecko – Co‑Design with Local Communities

• In Africa, Microsoft is piloting Project Gecko, co‑designing AI applications for agriculture and education in East Africa and South Asia, ensuring that safety criteria reflect local needs.

2.4 Collaboration Across the Global Evaluation Community

• Natasha highlighted existing collaborative efforts:
  • IMDA (Singapore) – brings together language and cultural experts from nine Asia‑Pacific nations for regional red‑teaming.
  • Sarai at IIT Madras (India) – produces culturally grounded testing that surfaces locally relevant bias and stereotyping.
  • GPAC Center (Tokyo) – proposes a multicultural AI consortium to move beyond translated prompts toward shared evaluation infrastructure for low‑resource languages.

3. Transition to Panel Discussion

• After the introductory remarks, Sunayana Sitaram took the stage as moderator.
• She introduced the panelists (representing industry, academia, research, and startups) and explained that each would receive a focused question to surface concrete experiences and ideas.

4. Panelist Contributions

4.1 Sara Hooker – The Most Urgent Safety Gaps

• Contextual Gap – Safety is culture‑specific; what is unsafe in the U.S. may be benign elsewhere.
• Illustrative Example – A widely‑used safety dataset from a frontier AI lab contained an American‑centric scenario (“threaten to go TP a house”). The phrase “TP” (to throw toilet‑paper) was unintelligible to many non‑US speakers, yet the dataset was used globally for alignment.
• Two‑fold Problem
  1. Talent Ecosystem Deficit – Need more local experts who can define harms in culturally relevant ways.
  2. Beyond Language Coverage – Safety must consider contextual nuances (e.g., social norms, indirect threats).
• Emerging Adversarial Issues – New harms arise as AI becomes cheap and ubiquitous:
  • Voice‑assistant fraud, targeted manipulation of vulnerable populations, and other socially engineered attacks.
• Adaptable Intelligence’s Goal – Build efficient, low‑cost adaptation pipelines that let a single model be quickly re‑calibrated for local safety expectations, rather than shipping a one‑size‑fits‑all set of red‑lines.

4.2 Wan Sie Lee – Lessons from Singapore’s IMDA Multilingual Red‑Teaming

• Translation is Not Sufficient – Simply translating a dataset does not guarantee comparable safety performance.
• Tool‑Calling Challenges – In agentic (tool‑using) scenarios, the wrong tool is often selected when prompts are translated; even the tool names may need localization.
• Standardization Difficulties –
  1. Data & Infrastructure – Multilingual testing requires bespoke datasets, evaluation pipelines, and toolkits that respect linguistic nuances.
  2. Cultural Expectations – What is deemed acceptable varies widely; building a common taxonomy that accommodates divergent cultural norms is a core challenge.
• Practical Takeaways
  • Need a shared understanding of what constitutes harm in each locale.
  • Development of standardized yet flexible taxonomies that can be extended for local contexts.

4.3 Additional Panel Voices (Not Fully Captured)

• Although the transcript excerpt does not contain explicit remarks from Fabrice Ciais (G42), Marzieh Fadaee (Cohere), Nicolas Miailhe (PrismEval), Nitarshan Rajkumar (Anthropic), and Wassim Hamidouche (Microsoft), the moderator noted that all were present. Their contributions are therefore [unidentified] in the available text.

5. Emerging Themes & Open Questions

Theme	Points of Consensus	Points of Divergence / Open Questions
Need for Local Talent	All panelists agree that local expertise is essential for defining culturally specific harms.	How to sustainably fund and scale such talent pipelines?
Evaluation Infrastructure	Shared benchmarks (AI Luminate, Samakisha) are seen as crucial.	What governance model will ensure open, reproducible, and unbiased evaluation across jurisdictions?
Standardization vs. Flexibility	Consensus that a common taxonomy is needed, but must be adaptable.	How to reconcile conflicting cultural expectations (e.g., freedom of expression vs. community protection)?
Adversarial Threats in Low‑Resource Settings	Recognized as a growing risk, especially with tool‑calling agents.	What concrete mitigation strategies (e.g., language‑specific guardrails) can be deployed at scale?
Efficient Model Adaptation	Sara Hooker’s venture aims for cheap, rapid localization.	What technical mechanisms (parameter efficient fine‑tuning, prompting, reinforcement learning) are most promising for multilingual safety?

6. Closing Remarks

• Natasha Crampton reiterated that trustworthy AI diffusion does not happen by accident; it requires deliberate design, evaluation, and governance that reflect the world’s linguistic and cultural diversity.
• The moderator thanked the panelists and signaled the transition to the next agenda item (not captured in the transcript).

Key Takeaways

• Safety mechanisms calibrated only on English miss culturally specific cues (idioms, indirect phrasing, humour), leading to higher failure rates in other languages.
• Multilingual safety gaps are both linguistic and contextual; the same behavior may be safe in one culture and harmful in another.
• Current benchmarks are expanding (AI Luminate, Samakisha) but require community‑driven data and evaluation pipelines to capture local nuances.
• Tool‑calling agents expose new multilingual challenges—translation alone cannot ensure correct tool selection or safe execution.
• A common, extensible taxonomy of harms is essential for cross‑country collaboration, yet must accommodate divergent cultural norms.
• Emerging adversarial threats (voice‑assistant fraud, AI‑driven scams) are amplified in low‑resource language contexts where safeguards are weakest.
• Efficient, low‑cost adaptation of safety judgments is the core mission of Adaptable Intelligence, aiming to let a single model serve many locales safely.
• Collaboration across governments, academia, and industry (IMDA, Sarai, GPAC, Microsoft, G42, Cohere, Anthropic, etc.) is already underway and must be deepened to build reusable, rigorous multilingual evaluation infrastructure.
• Future work must focus on building talent ecosystems, shared benchmarks, and governance frameworks that respect local values while maintaining global safety standards.

---

See Also:

• open-source-tooling-for-safe-secure-and-trustworthy-ai
• power-protection-and-progress-legislators-and-the-ai-era
• a-billion-voices-one-ai-how-language-tech-transforms-nations
• evaluations-and-open-source-software-for-ai-for-social-good-at-scale
• navigating-the-ai-regulatory-landscape-a-cross-compliance-framework-for-safety-and-governance
• ai-and-the-future-of-work-employability-skills-and-labour-market-transformation
• democratizing-ai-requires-a-free-and-open-internet
• responsible-ai-for-bharat-building-trust-safety-and-global-leadership
• ai-diplomacy-asia-leads-governing-artificial-intelligence-in-a-fragmented-world
• responsible-ai-at-scale-governance-integrity-and-cyber-readiness-for-a-changing-world